Agents enter the browser

AI agents are moving into the browser itself, turning Chrome, Edge, and similar software into the place where automated work gets deployed and supervised. (blog.google) Google said in a September 18, 2025 post that Gemini in Chrome can read the current page, work across tabs, and is adding “agentic browsing” so Chrome can handle tasks such as booking appointments. Google also said the feature was rolling out to desktop users in the United States and later to Google Workspace customers with admin controls. (blog.google) (workspaceupdates.googleblog.com) Microsoft made the same bet on July 28, 2025, when it introduced Copilot Mode in Edge as an experimental feature that can use multi-tab context and, with permission, help users compare pages and complete web tasks. Microsoft’s product page now describes Edge as “your AI browser.” (blogs.windows.com) (microsoft.com) An AI agent is software that does multi-step work, not just question answering; in browser products, that means reading pages, clicking buttons, filling forms, and carrying state from one tab to the next. OpenAI described that model on January 23, 2025 with Operator, a research preview that used its own browser, and later folded those capabilities into ChatGPT agent on July 17, 2025. (openai.com 1) (openai.com 2) The shift in 2025 and 2026 is that browser control is being packaged as infrastructure instead of a separate robot app. Cloudflare said on April 13, 2026 that it was expanding Agent Cloud to move agents from “local laptops” to “production-grade workloads,” and on April 15 it renamed Browser Rendering to Browser Run and called it “the browser for AI agents.” (cloudflare.com) (blog.cloudflare.com) That changes who manages the software. Google said Gemini in Chrome for Workspace comes with enterprise data protections and admin settings, while Microsoft said Copilot Mode requires permission for multi-tab access, putting control in browser policies and workplace accounts rather than in a standalone consumer tool. (workspace.google.com) (blogs.windows.com) The new design language around these systems reflects that move from chatbot to operating layer. Anthropic’s documentation for its computer use tool describes screenshot-based mouse and keyboard control for autonomous desktop interaction, which is closer to remote-control software than to a search box. (platform.claude.com) A parallel market is forming underneath the big browser brands. Browserbase says its service lets developers spin up browser sessions at scale for agents that handle forms, logins, uploads, and monitoring, while Cloudflare is adding live views, replay, and human intervention tools for the same class of workload. (browserbase.com) (blog.cloudflare.com) The result is that the browser is being recast as a managed work surface for artificial intelligence. The companies pushing that model are selling not just answers, but permissioned access to tabs, sessions, and business workflows. (microsoft.com) (blog.google)